according to the paper they get 98% accuracy. another recent paper came out saying it's always possible to discriminate between real and synthetic text [1].
i think the core problem is with the generalist classifiers (gptzero, openai detector, etc). ex. openai's classifier has an accuracy of around 25% on it's own text. however, when you train a bespoke classifier (like the authors did), you can get really good results.
Adversarial training isn't infinitely scalable either, has its limitations also.
Also - the moment that companies start training models to resist detectors, they expose themselves to regulation. Won't stop dark AI models running on some website somewhere, but it can be very effectively applied to companies running at Google or OpenAI scale.
i would recommend u read the paper. the contribution isnt a detector thats meant to be taken seriously; but a detector that works in a very specific task. they then use this to estimate use of LLMs on MTurk
